Every answer you get from Claude, ChatGPT or Le Chat runs on a chip in a data center that somebody had to finance, power and cool. In 2026 that has become the biggest industrial build-out of the decade: about $725 billion of capital spending by four US tech companies in a single year, quarterly Nvidia revenue close to $100 billion, and AI labs signing contracts measured not in servers but in gigawatts. This guide shows who has secured how much compute, on which chips, what Europe and Croatia are building, and what it means for companies that simply want reliable, affordable AI.
Big Tech is spending like never before
The four largest US cloud and platform companies plan to invest roughly $725 billion in 2026, up from a record $410 billion in 2025. Most of it goes into GPU clusters, custom chips and new data center buildings. After the first-quarter earnings calls, analysts at Evercore and Bank of America projected that the total will pass $1 trillion in 2027.
Amazon leads with about $200 billion, followed by Microsoft at about $190 billion. Part of the increase is not more capacity but higher prices: Microsoft says $25 billion of its 2026 budget is due to higher component pricing, and Meta raised its forecast partly for the same reason. Memory and networking gear have become scarce, which also matters for any company planning its own AI servers.
Nvidia: the clearest scoreboard
Most of that money ends up with one supplier. In its quarter ending 26 July 2026, Nvidia reported $96.2 billion of revenue, of which $89.0 billion came from data centers, more than double the year before. Data centers now make up about 92% of the company's sales. For the following quarter Nvidia guided to about $108 billion, without counting any data-center sales in China.
The labs: who has secured how much compute
The AI labs no longer buy servers. They sign multi-year contracts for power and chips measured in gigawatts. One gigawatt is roughly the output of a nuclear reactor; Mistral's CEO Arthur Mensch puts the cost of a gigawatt of AI compute at about $50 billion.
Anthropic has deliberately spread Claude across three chip families. In April 2026 it agreed up to 5 GW with Amazon, covering Trainium2, Trainium3 and Trainium4; over one million Trainium2 chips already run in Project Rainier, and close to 1 GW of Trainium capacity is due by the end of 2026. In the same month it expanded its deal with Google and Broadcom: about 1 GW of TPU capacity in 2026 and about 3.5 GW more from 2027. A further agreement with Microsoft and Nvidia adds up to 1 GW on Azure, and Anthropic has committed $50 billion to its own data centers in the US. The demand is real: annualised revenue rose from about $9 billion at the end of 2025 to over $30 billion by April 2026, with more than 1,000 customers each spending over $1 million a year.
OpenAI has gone even bigger. It has signed for about 26 GW of chips: 10 GW of Nvidia systems, 6 GW of AMD GPUs and 10 GW of custom accelerators designed with Broadcom. Its Stargate programme with Oracle and SoftBank reached 10 GW of contracted capacity years ahead of its 2029 target; the flagship site in Abilene, Texas runs at an estimated 0.6 GW. Estimates put OpenAI's total hardware and cloud commitments for 2025–2035 at about $1.15 trillion.
Mistral plays a different game: smaller, but its own and European. Manus, bought by Meta for more than $2 billion at the end of 2025, no longer needs to solve compute itself; its agents now run on the infrastructure of a company planning $115–135 billion of capex this year.
Europe and Croatia: building their own capacity
Mistral financed its first own data center with $830 million of debt in March 2026. The site at Bruyères-le-Châtel, about 30 km south of Paris, starts in the second half of 2026 with 13,800 Nvidia GB300 GPUs and 44 MW, and can grow to 1.4 GW. Mistral targets 200 MW by the end of 2027 and 1 GW by the end of 2029. In July 2026 the European Commission opened a €30 billion call for up to seven AI gigafactories of around 100,000 chips each; bids close on 12 November 2026.
Croatia is in the race too. The Pantheon AI campus in Topusko, about 45 minutes from Zagreb, is planned as a €50 billion project with 1 GW of capacity and 800 MW of IT load, powered by renewables including a 500 MW on-site solar plant. Croatia's grid operator HOPS approved the connection blueprint in July 2026; construction is due to start in early 2027, with full operation in 2029.
What this means for your company
- Capacity, not models, is the bottleneck. Rate limits, slower responses at peak hours and waiting lists for new models all come from scarce chips. For production systems, plan a fallback model and a second provider; see our guide to model routing.
- Use the cloud you already have. Because Anthropic runs on Trainium, TPU and Nvidia, Claude is available on AWS Bedrock, Google Vertex AI and Microsoft Foundry. OpenAI models run on Azure, Mistral on its own platform and all major clouds. Buying through your existing cloud keeps data, contracts and billing in one place.
- Token prices fall, total spend rises. More capacity and new chip generations keep pushing the price per token down, while agents use far more tokens per task. Budget by task, not by token; our inference cost guide shows how.
- Think twice about your own GPU servers. Component prices are rising, and the labs get priority from chipmakers. For most SMEs, renting capacity or using APIs is cheaper than buying hardware that will be a generation behind in 18 months.
- Energy is part of the story. AI data centers need gigawatts of power, and grid connections are now a gating factor. Our AI energy guide covers how much power Claude, ChatGPT and Mistral really use.
Bubble or foundation?
Not all of these gigawatts will be built on time, and some deals are circular: chipmakers invest in the labs that then buy their chips. OpenAI's commitments are many times its current revenue. But the demand behind them is measurable. Anthropic's revenue more than tripled within a few months, and Nvidia keeps beating forecasts. For a company using AI, the practical conclusion is the same in either scenario: do not tie yourself to one model or one cloud, and build processes that still work when you switch the engine underneath.
The bottom line
The compute race decides which models arrive first, how fast they respond and what they cost. Anthropic and OpenAI have secured tens of gigawatts each, Big Tech will spend more than $700 billion this year, and Europe, including a 1 GW project in Croatia, is starting to build its own capacity. For businesses, the winning strategy is not to pick the lab with the most chips, but to stay flexible: several models, a clear view of cost per task and data in the right jurisdiction.
Planning AI for your company?
We help companies in Croatia and the DACH region choose models, clouds and deployment options that fit their data, budget and compliance needs, from the first workshop to production agents. As a Claude Certified Architect based in Zagreb, we work with Claude, ChatGPT and Mistral every day.
Talk to an AI consultant